It is incompetence. Everytime I optimize any oss library developed by some Googlers, the result is 20X performance boost still using plain Python. Compiling my optimized version by Cython can only gain 1.5X speedup.

Like, comparing my Numpy code to Numpy code written by other people, I don't know why mine can be so much faster. (https://github.com/hirasawakinko/Numpy-Done-Right)